MFCC Features Research Articles

The Effectiveness of Unsupervised Subword Modeling With Autoregressive and Cross-Lingual Phone-Aware Networks

This study addresses unsupervised subword modeling, i.e., learning acoustic feature representations that can distinguish between subword units of a language. We propose a two-stage learning framework that combines self-supervised learning and cross-lingual knowledge transfer. The framework consists of autoregressive predictive coding (APC) as the front-end and a cross-lingual deep neural network (DNN) as the back-end. Experiments on the ABX subword discriminability task conducted with the Libri-light and ZeroSpeech 2017 databases showed that our approach is competitive or superior to state-of-the-art studies. Comprehensive and systematic analyses at the phoneme- and articulatory feature (AF)-level showed that our approach was better at capturing diphthong than monophthong vowel information, while also differences in the amount of information captured for different types of consonants were observed. Moreover, a positive correlation was found between the effectiveness of the back-end in capturing a phoneme's information and the quality of the cross-lingual phone labels assigned to the phoneme. The AF-level analysis together with t-SNE visualization results showed that the proposed approach is better than MFCC and APC features in capturing manner and place of articulation information, vowel height, and backness information. Taken together, the analyses showed that the two stages in our approach are both effective in capturing phoneme and AF information. Nevertheless, monophthong vowel information is less well captured than consonant information, which suggests that future research should focus on improving capturing monophthong vowel information.

Phonocardiogram Signal Based Multi-Class Cardiac Diagnostic Decision Support System

A Phonocardiogram (PCG) signal represents murmurs and sounds signals made by vibrations caused for the period of a cardiac cycle. Acoustic wave generated through the beat of the cardiac cycle propagates through the chest wall. It can be easily recorded by a low-cost small handheld digital device called a stethoscope. It provides information like heart rate, intensity, tone, quality, frequency, and location of various components of cardiac sound. Due to these characteristics, phonocardiogram signals can be used to detect heart status at an early stage in a non-invasive manner. In previous studies, the Convolutional Neural Network (ConvNet) is the most studied architecture, which was fed by features, namely Mel Frequency Cepstral (MFC), Chroma Energy Normalized Statistics (CENS), and Constant-Q Transform (CQT). This work has proposed a ConvNet model trained by Hybrid Constant-Q Transform (HCQT) for heart sound beat classification. CQT, Variable-Q Transform (VQT), and HCQT are extracted from each phonocardiogram signal as the acoustic features, including the dominant MFCC features, feed into five-layer regularized ConvNets. After analyzing the literature in the same domain, it can be stated that this is the first time HCQT is being utilized for PCG signals. The findings of the experiments demonstrate that HCQT is more effective than standard CQT and other variants. Also, the accuracies of the system proposed in this work on the validation datasets are 96% in multi-class classification, which outperforms the proposed work relative to other models significantly. The source code is available on the Github repository https://github.com/shamiktiwari/ PCG-signal-Classification-using-Hybrid-Constant-Q-Transform to support the research community.

Exploring the mel scale features using supervised learning classifiers for emotion classification

Human emotions are inherently ambiguous and impure but emotions are important while considering the human uttered speech. The role of human speech is intensified by the aspect of the emotion it conveys. There are several characteristics of speech that differentiates it among different utterances. Various prosodic features like pitch, timbre, loudness and vocal tone categorise speech into several emotions and other domains. The sample speech is changed when it is subjected to various emotional environments. Researches support various experimental analyses for phonetics and prosodic parameters that quantify the quality of speech. Identification of different emotional states of an actor (speaker) can also be done on the basis of mel scale. MFCC is one such variant to study the emotional aspects of the utterances by the speaker. The paper implements a model to identify several emotional states from MFCC for two datasets. The work classifies emotions for two datasets on the basis of MFCC features and gives the comparison of both. This work implements a classification model based on dataset minimisation that is done by taking the mean of features for the improvement of classification accuracy on different machine learning algorithms.

Precision Adaptive MFCC Based on R2SDF-FFT and Approximate Computing for Low-Power Speech Keywords Recognition

This paper presents system-architecture-circuits co-designs for computing the MFCC feature extraction for speech keywords recognition. The trade-off between accuracy and power consumption under various background noises is achieved by using the 8-stage radix-2 single-path delay feedback FFT (R2SDF-FFT) and the precision self-adaptive architecture with approximate computing. The R2SDF-FFT structure with the fine-grained bit-width quantization can reduce 35.7% of memory size. Approximate multiplication and addition with Dual-Vdd are proposed to further improve the FFT computing energy efficiency. Finally, we present the precision self-adaptive MFCC architecture with the proposed FFT, which can be dynamically configured to use two calculation modes with different hardware settings according to the input speech background noise. Implemented and evaluated under 22 nm technology, the power consumption of the proposed design can be reduced by up to 76.3%, while the accuracy increased by 0.8%.

Lung disease classification using machine learning algorithms

In this study we compared support vector machines (SVM), k-nearest neighbor (k-NN), and Gaussian Bayes (GB) algorithms in classification of respiratory diseases with text and audio data. An electronic stethoscope and its software are used to record patient information and 17930 lung sounds from 1630 subjects. SVM, k-NN and GB algorithms were run on 6 datasets to classify patients into; (1) sick or healthy with text data, (2) sick or healthy with audio MFCC features, (3) sick or healthy with the text data and audio MFCC features, (4) 12 diseases with text data, (5) for 12 disease with audio MFCC features, (6) for 12 disease with the text data and audio MFCC features. Accuracy results in SVM were %75, %88, %64, %73, %63, %70; for k-NN %95, %92, %92, %67, %64, %66; for GB %98, %91, %97, %58, %48, %58 respectively. In 12 class classification of lung diseases, the most accurate algorithm was SVM with text data. In classifying via audio data, k-NN was the most accurate. Using both audio and text data, SVM was the most accurate. When we classify healthy versus sick via text, audio and combined data, GB was always the most accurate with very high accuracy, closely followed by k-NN. We can infer from here that when we have large number of features but limited amount of samples, SVM and k-NN are best in classifying the dataset in more than two classes. However GB is best when it comes to classifying into two classes.

Big Data Classification for the Analysis MEL Scale Features Using KNN Parameterization

The role of human speech is intensified by the emotion it conveys. The parameterization of the vector obtained from the sentence divided into the containing emotional-informational part and the informational part is effectively applied. There are several characteristics and features of speech that differentiate it among utterances, i.e. various prosodic features like pitch, timbre, loudness and vocal tone which categorize speech into several emotions. They were supplemented by us with a new classification feature of speech, which consists in dividing a sentence into an emotionally loaded part of the sentence and a part that carries only informational load. Therefore, the sample speech is changed when it is subjected to various emotional environments. As the identification of the speaker’s emotional states can be done based on the Mel scale, MFCC is one such variant to study the emotional aspects of a speaker’s utterances. In this work, we implement a model to identify several emotional states from MFCC for two datasets, classify emotions for them on the basis of MFCC features and give the comparison of both. Overall, this work implements the classification model based on dataset minimization that is done by taking the mean of features for the improvement of the classification accuracy rate in different machine learning algorithms.

Detection of Snore from OSAHS Patients Based on Deep Learning.

Obstructive sleep apnea-hypopnea syndrome (OSAHS) is extremely harmful to the human body and may cause neurological dysfunction and endocrine dysfunction, resulting in damage to multiple organs and multiple systems throughout the body and negatively affecting the cardiovascular, kidney, and mental systems. Clinically, doctors usually use standard PSG (Polysomnography) to assist diagnosis. PSG determines whether a person has apnea syndrome with multidimensional data such as brain waves, heart rate, and blood oxygen saturation. In this paper, we have presented a method of recognizing OSAHS, which is convenient for patients to monitor themselves in daily life to avoid delayed treatment. Firstly, we theoretically analyzed the difference between the snoring sounds of normal people and OSAHS patients in the time and frequency domains. Secondly, the snoring sounds related to apnea events and the nonapnea related snoring sounds were classified by deep learning, and then, the severity of OSAHS symptoms had been recognized. In the algorithm proposed in this paper, the snoring data features are extracted through the three feature extraction methods, which are MFCC, LPCC, and LPMFCC. Moreover, we adopted CNN and LSTM for classification. The experimental results show that the MFCC feature extraction method and the LSTM model have the highest accuracy rate which was 87% when it is adopted for binary-classification of snoring data. Moreover, the AHI value of the patient can be obtained by the algorithm system which can determine the severity degree of OSAHS.

Performance of a language identification system using hybrid features and ANN learning algorithms

This paper discusses the use of hybrid features and artificial neural network (ANN) for spoken language identification (LID). At the feature extraction stage, we have used RASTA-PLP features and later hybrid features by combining the state-of-the-art MFCC features with RASTA-PLP features. The classifier used at the back-end of the LID system is Feed-forward Back Propagation Neural Network (FFBPNN). Performance comparison is done for all the thirteen learning algorithms available in FFBPNN. Results indicate better performance with MFCC + RASTA-PLP features as compared to RASTA-PLP features when used individually. The classification accuracy of 94.6 percent is obtained with ‘trainlm’ network training function and a test error rate of 0.10 with MFCC + RASTA-PLP hybrid features. On the other hand, RASTA-PLP features provide the best classification accuracy of 89.6%. Also, in this paper two new training functions are proposed. Experimental results show that for the MFCC + RASTA-PLP features, the proposed training function 1 offers 95.1 percent accuracy and 0.0993 mean square error. Similarly, the proposed training function 2 offers 95.3 per cent classification accuracy and 0.093 minimum mean square error for MFCC + RASTA-PLP features. The feasibility of the proposed solution is estimated by simulating multiple experiments on a language database of four languages i.e. Hindi, Tamil, Malayalam and English in the working Platform of MATLAB.

A new fusion feature based on convolutional neural network for pig cough recognition in field situations

Pig cough is considered the most common clinical symptom of respiratory diseases. Thus, establishing an early warning system for respiratory diseases in pigs by monitoring and identifying their cough sounds is important. In this paper, we propose a new fusion feature, namely Mel-frequency cepstral coefficient-convolutional neural network (MFCC-CNN), to improve the recognition accuracy of pig coughs. We obtained the MFCC-CNN feature by fusing multiple frames of MFCC with multiple one-layer CNNs. We used softmax and linear support vector machine (SVM) classifiers for classification. We tested the algorithm through field experiments. The results reveal that the performance of classifiers using the MFCC-CNN feature was significantly better than those using the MFCC feature. The F1-score increased by 10.37% and 5.21%, and the cough accuracy increased by 7.21% and 3.86% for the softmax and SVM classifiers, respectively. We also analyzed the impact of different numbers of fusion frames on the classification performance. The results reveal that fusing 55 and 45 adjacent frames resulted in the best performance for the softmax and SVM classifiers, respectively. From this research, we can conclude that a system constructed by simple one-layer CNNs and SVM classifiers can demonstrate excellent performance in pig sound recognition.

Recognition of Isolated Digit Using Random Forest for Audio-Visual Speech Recognition

The proposed research work clearly investigates the effective use of two modalities (audio and visual inputs) toward designing functional audio-visual speech recognition system. The promising results presented in this piece of work were obtained on vVISWa (visual Vocabulary of Isolated Standard Words) dataset of audio-visual words and CUAVE (Clemson University Audio-Visual Experiments) database, respectively. The discrete cosine transform (DCT), local binary pattern (LBP) features of full frontal visual profile and MFCC features for acoustics signals were fused together for recognition purpose and were classified using random forest classifier.

Phonocardiogram Classification Based on Machine Learning with Multiple Sound Features

In this study the heartbeat sound signals were tackled by classifying them into heart disease categories such as normal, artifact, murmur and extrahals in an attempt for early detection of heart defects. Phonocardiogram (i.e., PCG) is used to obtain the digital recording dataset of the heart sounds using an electronic stethoscope or mobile device. Multiple features are extracted from the digital recording dataset such as MFCC, Delta MFCC, FBANK and a combination between MFCC and FBANK features. Moreover, to classify the heartbeat sound signals, multiple well-known machine learning classifiers were used such as Naive Bays (NB), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), K-Nearest Neighbor (KNN) and Artificial Neural Network (ANN). The evaluation processes went through five metrics: Confusion matrix, accuracy, F1 score, precision and recall evaluating the recognition rate. Comparative experimental results show that the correctness of the feature with a best accuracy 99.2% adopted by MFCC and FBANK combination features which reduce false detection.

Bionic optimization of MFCC features based on speaker fast recognition

Surrounded by low SNR, how to make the voice faster and better recognize the owner has become a heated research topic. The human auditory system can accurately acquire the characteristics of acoustic events in complex systems or low SNR noise environment, which is of significance in the research of bionic hearing of human ear. The response curve of human ear output is obtained by bionic technology, which is the best response curve for sound enhancement to modify Mel filter. The method of adaptive threshold selection is used to integrate Mel features to realize the reduction and dynamic extraction of low SNR speech features. This method not only can resist the disadvantages of poor robustness and complexity of parameter model, but also obtain dynamic and comprehensive speech information of different speakers in different scenes. Finally, the improved CNN and I-vector system are contributed to reduce the dimension of the data and to verify the recognition, so as to achieve the optimal frequency selective amplification and simplification of the acoustic signal. In the case of SNR-5db, the model is reduced by 15% and the recognition accuracy is improved by 3%.

Prosody-Based Measures for Automatic Severity Assessment of Dysarthric Speech

One of the first cues for many neurological disorders are impairments in speech. The traditional method of diagnosing speech disorders such as dysarthria involves a perceptual evaluation from a trained speech therapist. However, this approach is known to be difficult to use for assessing speech impairments due to the subjective nature of the task. As prosodic impairments are one of the earliest cues of dysarthria, the current study presents an automatic method of assessing dysarthria in a range of severity levels using prosody-based measures. We extract prosodic measures related to pitch, speech rate, and rhythm from speakers with dysarthria and healthy controls in English and Korean datasets, despite the fact that these two languages differ in terms of prosodic characteristics. These prosody-based measures are then used as inputs to random forest, support vector machine and neural network classifiers to automatically assess different severity levels of dysarthria. Compared to baseline MFCC features, 18.13% and 11.22% relative accuracy improvement are achieved for English and Korean datasets, respectively, when including prosody-based features. Furthermore, most improvements are obtained with a better classification of mild dysarthric utterances: a recall improvement from 42.42% to 83.34% for English speakers with mild dysarthria and a recall improvement from 36.73% to 80.00% for Korean speakers with mild dysarthria.

A low latency modular-level deeply integrated MFCC feature extraction architecture for speech recognition

In this paper, a low-complex chip to extract the Mel Frequency Cepstral Coefficient for a speech recognition system is presented. The architecture can operate in a continuous-flow manner to process streaming or the stored speech signal at high speed. The frame-overlap Hamming window, DFT and Mel-filter bank computations are deeply integrated to share memory buffers and avoid bit-reversal circuit to reduce area and latency. Moreover, normalised energy consumption and area delay product are reduced by 32%, and speed is increased by 5.2 times compared to prior works. Further, the fixed-point word-length is optimised to minimise the area without affecting the accuracy.

Pattern recognition and features selection for speech emotion recognition model using deep learning

Automatic speaker recognizing models consists of a foundation on building various models of speaker characterization, pattern analyzing and engineering. The effect of classification and feature selection methods for the speech emotion recognition is focused. The process of selecting the exact parameter in arrangement with the classifier is an important part of minimizing the difficulty of system computing. This process becomes essential particularly for the models which undergo deployment in real time scenario. In this paper, a new deep learning speech based recognition model is presented for automatically recognizes the speech words. The superiority of an input source, i.e. speech sound in this state has straight impact on a classifier correctness attaining process. The Berlin database consist around 500 demonstrations to media persons that is both male and female. On the applied dataset, the presented model achieves a maximum accuracy of 94.21%, 83.54%, 83.65% and 78.13% under MFCC, prosodic, LSP and LPC features. The presented model offered better recognition performance over the other methods.

Comparison of CNNs and SVM for voice control wheelchair

In this paper, we develop an intelligent wheelchair using CNNs and SVM voice recognition methods. The data is collected from Google and some of them are self-recorded. There are four types of data to be recognized which are go, left, right, and stop. Voice data are extracted using MFCC feature extraction technique. CNNs and SVM are then used to classify and recognize the voice data. The motor driver is embedded in Raspberry PI 3B+ to control the movement of the wheelchair prototype. CNNs produced higher accuracy i.e. 95.30% compared to SVM which is only 72.39%. On the other hand, SVM only took 8.21 seconds while CNNs took 250.03 seconds to execute. Therefore, CNNs produce better result because noise are filtered in the feature extraction layer before classified in the classification layer. However, CNNs took longer time due to the complexity of the networks and the less complexity implementation in SVM give shorter processing time.

Recognize basic emotional statesin speech by machine learning techniques using mel-frequency cepstral coefficient features

Speech Emotion Recognition (SER) has been widely used in many fields, such as smart home assistants commonly found in the market. Smart home assistants that could detect the user’s emotion would improve the communication between a user and the assistant enabling the assistant to offer more productive feedback. Thus, the aim of this work is to analyze emotional states in speech and propose a suitable algorithm considering performance verses complexity for deployment in smart home devices. The four emotional speech sets were selected from the Berlin Emotional Database (EMO-DB) as experimental data, 26 MFCC features were extracted from each type of emotional speech to identify the emotions of happiness, anger, sadness and neutrality. Then, speaker-independent experiments for our Speech emotion Recognition (SER) were conducted by using the Back Propagation Neural Network (BPNN), Extreme Learning Machine (ELM), Probabilistic Neural Network (PNN) and Support Vector Machine (SVM). Synthesizing the recognition accuracy and processing time, this work shows that the performance of SVM was the best among the four methods as a good candidate to be deployed for SER in smart home devices. SVM achieved an overall accuracy of 92.4% while offering low computational requirements when training and testing. We conclude that the MFCC features and the SVM classification models used in speaker-independent experiments are highly effective in the automatic prediction of emotion.

Evaluation and modeling of automotive transmission whine noise quality based on MFCC and CNN

The sound quality of automotive transmission noise strongly influences passengers’ psychological and physiological perceptions. To predict the sound quality of automotive transmission noise, a uniform deceleration noise test with two automotive transmissions has been conducted in a semi-anechoic room. All recorded transmission noise signals have been divided into 5 s segments and subsequently evaluated subjectively through the rating scales test by a jury. In addition, a novel prediction method, namely, Mel-Frequency Cepstral Coefficients-based convolutional neural networks (MFCC-CNN), which substitute the softmax classification layer for the linear transform prediction layer at the output of the general CNN’s structure and take MFCC feature map as input, has been proposed to predict the transmission sound quality. MFCC's distinguishing performance on sound quality has been validated. The parameter selection of the MFCC-CNN model has been compared and studied using a grid search. In addition, three conventional machine-learning-based methods have been introduced to enable a comparison of the performance with the newly developed MFCC-CNN. The results show the following: (1) In different transmission gears, MFCC features can distinguish different sound quality noises. (2) The accuracy of the proposed MFCC-CNN sound quality prediction approach are better than those of the 3 other referenced methods. (3) The correlation coefficient of prediction value from MFCC-CNN is more than 0.95 and the mean absolute error of prediction value from MFCC-CNN is less than 0.55, which fully meets the need of engineering. Finally, the newly proposed MFCC-CNN approach may be extended to address other vehicle noises in the future.

A Novel automatic voice recognition system based on text-independent in a noisy environment

Automatic voice recognition system aims to limit fraudulent access to sensitive areas as labs. Our primary objective of this paper is to increase the accuracy of the voice recognition in noisy environment of the Microsoft Research (MSR) identity toolbox. The proposed system enabled the user to speak into the microphone then it will match unknown voice with other human voices existing in the database using a statistical model, in order to grant or deny access to the system. The voice recognition was done in two steps: training and testing. During the training a Universal Background Model as well as a Gaussian Mixtures Model: GMM-UBM models are calculated based on different sentences pronounced by the human voice (s) used to record the training data. Then the testing of voice signal in noisy environment calculated the Log-Likelihood Ratio of the GMM-UBM models in order to classify user's voice. However, before testing noise and de-noise methods were applied, we investigated different MFCC features of the voice to determine the best feature possible as well as noise filter algorithm that subsequently improved the performance of the automatic voice recognition system.

Morlet-Kernel Principal Component Analysis Features for Speech Recognition

The aim of this paper is to present the application of Morlet wavelet to extract the speech features in place of MFCC features. KPCA is applied for selecting and reducing the large features obtained from Morlet wavelet. NLMS (Normalized Least Mean Square) filter is used to reduce additive noise levels ranging from ±5 dB to ±15 dB. Features are modeled using Ensembled Support Vector Machine classification model for FSDD and Kannada multi speaker data sets. The comparative results are discussed over logistic regression model. The proposed model reduces the noise with 99% of recognition rate for isolated words. The efficiency of ensembled classification model is explored.